Inter-sentence Context Modeling and Structure-aware Representation Enhancement for Conversational Sentiment Quadruple Extraction

Yu Zhang, Zhaoman Zhong, Huihui LvEMNLP 2025

TL;DR

This paper introduces ICMSR, a novel framework for extracting sentiment quadruples (target, aspect, opinion, sentiment) from conversations. It tackles the challenge of fragmented dialogue semantics by enhancing inter-sentence context modeling and leveraging dialogue structure, achieving state-of-the-art results on benchmark datasets.

What Problem Does This Paper Address?

Extracting fine-grained sentiment information from conversations is crucial for understanding user opinions. The task of Conversational Aspect-Based Sentiment Quadruple Analysis (DiaASQ) aims to extract quadruples consisting of target entities, aspects, opinions, and their associated sentiment polarities. Existing methods often struggle with this task due to two main limitations:

  1. Inadequate Inter-utterance Modeling: Traditional approaches often process utterances sequentially, leading to a loss of contextual coherence and difficulty in capturing long-range semantic dependencies across turns in a conversation.
  2. Underutilization of Dialogue Structure: Dialogue structures, such as speaker roles, topic threads, and reply relationships, are vital for understanding conversational dynamics. However, existing methods often employ static pooling strategies that fail to dynamically integrate these structural features, resulting in incomplete semantic representations.

Methods

The proposed ICMSR model addresses these limitations with two core modules:

  • Dialog Inter-sentence Contextual Enhancer (DICE): This module enhances inter-utterance interactions after initial sentence encoding. It uses a Transformer-like architecture with cross-attention mechanisms to explicitly model dependencies between sentences, thereby improving contextual continuity and mitigating fragmentation. DICE is applied in multiple layers to iteratively refine sentence representations.
  • Dialog Feature Amplifier (DFA): This module leverages dialogue structural features to enrich representations. It comprises two submodules:
  • STREAM (Speaker-Thread-Reply Enhanced Aggregation Module): This module dynamically fuses speaker roles, topic threads, and reply dependencies. It uses cross-attention to align these heterogeneous structural features and then fuses them with learnable weights. An Improved Enhancement Layer (IEL) further processes these fused features.
  • SMM (Structural Multi-Scale Mechanism): SMM enhances the structural representations by capturing hierarchical dependencies and multi-scale semantic patterns. It processes features through parallel layers with different expansion rates to model structure at varying resolutions, then uses global average pooling and learned weights to aggregate these multi-scale representations.

Datasets

The paper uses the DiaASQ dataset constructed by Li et al. (2023) as the foundation for their study. This dataset is sourced from Weibo and includes:

  • A Chinese version with 7452 discourses and 5742 quaternions.
  • An English version with 5514 quaternions.

The dataset is suitable for DiaASQ research due to its typical multi-turn interactions and dynamic changes.

Key Insights and Ideas

  • Inter-sentence context is key: Explicitly modeling dependencies between utterances, not just within them, is crucial for capturing the full semantic flow of a conversation. DICE's multi-layer Transformer-like approach effectively achieves this.
  • Structure matters more than static pooling: Simply pooling structural features (speaker, thread, reply) is insufficient. Dynamically fusing and refining these features, as DFA does, significantly improves performance.
  • Multi-scale structural analysis is beneficial: SMM's approach of processing structural information at different resolutions helps capture nuanced relationships, particularly important for complex dialogues.
  • Dialogue complexity influences optimal model depth: The optimal number of DICE layers varies between Chinese and English datasets, suggesting that the inherent complexity of dialogue structures (e.g., Chinese dialogues being more complex) impacts how deep inter-sentence modeling needs to be.

Key Results

The ICMSR model consistently achieves state-of-the-art performance on both Chinese (ZH) and English (EN) DiaASQ datasets across entity recognition, pairwise relation extraction (T-A, T-O, A-O), and the full sentiment quadruple extraction task.

  • On ZH: ICMSR surpasses IFusionQuad by significant margins on pairwise subtasks and improves Mi-F1 and Id-F1 by 1.02 and 0.64 respectively.
  • On EN: ICMSR achieves gains of 3.40 Mi-F1 and 2.57 Id-F1.
  • Compared to DiaASQ: ICMSR shows substantial improvements, with gains of +7.61/+7.69 on ZH and +6.05/+7.26 on EN for Mi-F1/Id-F1.
  • Ablation studies confirm the significant contributions of both DICE and DFA. Removing DICE leads to a noticeable drop in Mi-F1, especially on ZH, while removing DFA causes a substantial decline, particularly on EN. STREAM is more prominent for EN, while SMM is more beneficial for ZH.
  • Cross-utterance analysis shows ICMSR maintains strong performance even at longer utterance distances (Cross-3), outperforming baselines that fail to extract valid quadruples.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Limited Cross-Domain Validation: The model has only been validated on the DiaASQ dataset. Its generalization capability to other domains or platforms remains unverified, and further exploration is needed.
  • Dependency on Pre-trained Language Models (PLMs): The approach heavily relies on PLMs. Future work could focus on enhancing cross-lingual generalizability and developing robust solutions for resource-constrained settings or scenarios with significant data distribution shifts.
  • Insufficient Exploration of Multi-party Dialogue Scenarios: While effective for multi-turn dialogues, the model's effectiveness in more complex multi-party conversational settings is limited. Future work plans to incorporate additional features (e.g., syntactic information) to address these intricate interaction patterns.

Summary updated

BibTeX
@inproceedings{zhang2025inter,
  author = {Yu Zhang and Zhaoman Zhong and Huihui Lv},
  title = {{Inter-sentence Context Modeling and Structure-aware Representation Enhancement for Conversational Sentiment Quadruple Extraction}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure